Data Product (3) Can Data Actually Improve Real-World Operations?

The existing car wash operation only checked whether a vehicle was dirty before requesting a wash, which meant unnecessary requests could occur even when it was raining or snowing. Believing that factoring in precipitation forecasts could reduce costs while improving operational efficiency, the team quantitatively verified the relationship between weather and vehicle contamination.

The weather data was reorganized around the Korea Meteorological Administration's short-term forecasts. The frequent failures in loading external data were determined to stem from a lack of structure, since only fragments of necessary data had been accumulated piecemeal. After discussing with data users, the dataset was restructured to fit its intended purpose, and the data engineering team stabilized the loading process.

For the analysis, instead of using administrative district boundaries, the team chose to link vehicle locations with nationwide Automated Surface Observing System (ASOS) stations. To create a standard suited to actual operations, vehicles were matched to the nearest observation station, and weather and vehicles were linked within a 32km radius to reduce coverage gaps.

Looking at the correlation with vehicle contamination, rain and snow needed to be interpreted differently. Using thresholds of 3mm or more per hour for rainfall and 1cm or more per hour for snowfall, vehicles present within ±2 hours of that point in time were defined as having "been rained on" or "been snowed on."

As a result, vehicles that had been snowed on were more likely to be contaminated, so it was reasonable to maintain wash requests as before. On the other hand, when rainfall exceeded a certain level, vehicles in outdoor zones tended to have their exteriors naturally cleaned instead—leading to the conclusion that holding off on wash requests under those conditions would be more efficient.

Operational implementation was re-optimized based on forecast data. Using the short-term forecast updated at 5 AM, hourly precipitation amounts and probabilities for 6 AM to 11 PM that same day were applied, and the threshold values for hourly precipitation and probability were adjusted to find the balance point between actual prediction accuracy and operational effectiveness.

Simulations confirmed that a policy of postponing wash requests while it was raining and processing already-issued requests once the weather cleared reduced the number of wash requests without creating a problem of requests piling up at a specific point in time. This policy has since operated stably for over a year, with confirmed improvements in operational efficiency and cost savings.

The key takeaway is that intuitive judgment alone isn't enough—it must be converted into criteria that jointly account for precipitation amount, precipitation probability, regional differences, and time of day in order to be usable in actual operations. Data doesn't provide answers on its own; only when organized into context-reflecting rules does it become a tool that can actually change operational decision-making.